How should legal handle a claim that an AI image copies someone's work?

Pull the asset record first: inputs, references, model, approvals and placements. If a third-party image was used as a reference, consider taking the asset down while you assess; otherwise compare the works as you would in any similarity claim. Preserve the records, respond within the deadline in the letter, and check whether an indemnity requires prompt notice to the vendor.

Can we use photos of real people as references for AI generation?

Only with a release that covers it. A reference showing an identifiable person is personal data, and generating a likeness from it can raise image-rights or publicity claims as well as data protection issues. Use references of people only with signed consent that explicitly covers AI generation, and keep that consent with the asset record.

What does Synthetic White record for legal and compliance teams?

Every asset is logged with the model that produced it, and approvals follow defined roles, with each sign-off captured in an audit trail. Approved claims and banned words are enforced as content is generated, commercially safe models are the default, nothing customers upload trains any model, and model providers work under zero-retention agreements.

AI Content for Legal and Compliance Teams

AI content creation for legal and compliance teams.

Legal and compliance teams can approve AI content creation once four things are in place: a record of how every asset was made, models cleared for commercial use in writing, claims and disclaimers controlled while content is generated, and clear points where legal must sign off. With those, legal spends its time on judgement calls instead of reconstructing an asset's history after a complaint arrives.

The requests legal receives, and the records that answer them

Sooner or later legal will field an advertising complaint that needs substantiation, an infringement letter claiming an image copies someone's work, a subject access request from a person who appears in content, or a litigation hold. Each needs facts fast. For every asset, keep the final file and earlier versions, the brief and prompts, every reference input, the model that produced it, the approvals with names and dates, and where and when it ran. Reference inputs matter most in infringement disputes, because an uploaded photographer's image changes the analysis. Records should be exportable, and it must be possible to freeze them when a dispute is anticipated.

Commercially cleared, and in writing

A platform's terms are only the top layer; beneath them sit the terms of each model provider. Legal needs, in writing, the commercial-use licence for outputs, who holds any rights in them, what licence the vendor takes to your inputs, whether inputs train anything, which uses are restricted, and whether an IP indemnity exists, with its conditions and cap. Ask to be told before new model providers are added. Then set a rule inside the organisation: published work uses only cleared models. Synthetic White defaults to commercially safe models and logs the model behind every asset, so the rule can be checked rather than assumed.

Claims control at the point of generation

Generative copy tools reach for superlatives, health benefits and green adjectives, because that is what marketing text usually sounds like. Reviewing every draft for them is slow; stopping them at generation is faster. Hold approved claims, with their substantiation, as a library that copy must draw from, keep mandatory disclaimers as locked text per market, and maintain banned-word lists. Update them when the law moves. From 27 September 2026, EU rules under Directive 2024/825 prohibit generic environmental claims, such as 'green' or 'eco-friendly', unless the trader can show recognised excellent environmental performance, and claims that a product's impact is neutral or reduced because emissions are offset; add those terms to banned lists for EU markets and check how each country has implemented the rules.

Disclosure positions and non-negotiable sign-offs

Two decisions belong to legal. The first is the house position on disclosure: a matrix by market, channel and realism that says when AI use is labelled, with standard wording, so nobody improvises. EU deepfake disclosure under the AI Act applies from 2 August 2026, platforms add labels of their own, and France already requires the mention 'Images virtuelles', or an equivalent, on AI-generated faces or bodies in influencer promotions. The second is the list of content that always needs legal sign-off: new or comparative claims, regulated products, real people's likeness or voice, children, prize promotions and environmental claims. Everything else can follow brand and market approval, with a periodic sample checked by legal.

Updated 25 September 2026